ToolSiftToolSift
Back to tutorials

Dify deployment notes

Open-source LLM app platform for agents, workflows, and knowledge bases.

Deployment verdict

Dify is best read as an LLM application platform rather than a simple chatbot. Its value is strongest when a team wants a visible workflow layer for prompts, knowledge bases, agents, model routing, and app publishing. The tradeoff is operational complexity: once you move beyond a demo, credentials, datasets, plugin behavior, user permissions, and model cost control all need ownership.

Before installing

  • Review the license: Apache-2.0.
  • Check whether Docker is supported: yes.
  • Check API key dependency: depends on model/provider.
  • Confirm supported models: OpenAI, Claude, Gemini, Ollama, DeepSeek, Qwen.

Recommended deployment path

  1. Start with the official Docker Compose path in a disposable environment.
  2. Connect one model provider first; avoid enabling every provider at once.
  3. Create a small knowledge base with 20 to 50 known documents, then test answer quality against questions with known answers.
  4. Only after retrieval quality is acceptable should you add users, plugins, and external workflows.

Common evaluation traps

  • A successful login page does not prove the RAG pipeline is useful.
  • Plugin and tool permissions should be reviewed before connecting private systems.
  • The real cost is often model calls plus maintenance, not the open-source license.

Acceptance test tasks

  1. Build one knowledge-base app from a controlled document set.
  2. Ask five factual questions, three synthesis questions, and two adversarial questions.
  3. Export or document the prompt, model, retrieval settings, and failure cases.

Setup commands

git clone https://github.com/langgenius/dify.gitcd dify/dockercp .env.example .envdocker compose up -d